AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles (e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities, and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.

Tarski's writings on the concepts of truth and logical consequence rank among the most influential works in both logic and philosophy of the twentieth century. Because of this, it would be impossible to give a careful and accurate account of how far that influence reaches and of the complex route by which it spread. In logic, Tarski's methods of defining satisfaction and truth, as well as his work pioneering general model-theoretic techniques, have been entirely absorbed into the way the subject is presently done; they have become part of the fabric of contemporary logic, material presented in the initial pages of every modern textbook on the subject. In philosophy, the influence has been equally pervasive, extending not only to work in semantics and the philosophies of logic and language, but to less obviously allied areas such as epistemology and the philosophy of science as well.Rather than try to chart the wide-ranging influence of these writings or catalog the important research they have inspired, I will concentrate on various confusions and misunderstandings that continue to surround this work. For in spite of the extensive attention the work has received in the past fifty years, especially in the philosophical literature, misunderstandings of both conceptual and historical sorts are still remarkably widespread. Indeed in the philosophical community, recent reactions to Tarski's work on truth range from Karl Popper's “intense joy and relief” at Tarski's “legitimation” of the notion [1974, p. 399], to Hilary Putnam's assessment that “as a philosophical account of truth, Tarski's theory fails as badly as it is possible for an account to fail” [1985, p. 64]. Opinions have not exactly converged.

No abstract

No abstract

A major concern to the founders of modern logic—Frege, Peirce, Russell, and Hilbert—was to give an account of the logical structure of valid reasoning. Taking valid reasoning in mathematics as paradigmatic, these pioneers led the way in developing the accounts of logic which we teach today and that underwrite the work in model theory, proof theory, and definability theory. The resulting notions of proof, model, formal system, soundness, and completeness are things that no one claiming familiarity with logic can fail to understand, and they have also played an enormous role in the revolution known as computer science. The success of this model of inference led to an explosion of results and applications. But it also led most logicians—and those computer scientists most influenced by the logic tradition—to neglect forms of reasoning that did not fit well within this model. We are thinking, of course, of reasoning that uses devices like diagrams, graphs, charts, frames, nets, maps, and pictures. The attitude of the traditional logician to these forms of representation is evident in the quotation of Neil Tennant in Chapter I, which expresses the standard view of the role of diagrams in geometrical proofs. One aim of our work, as explained there, is to demonstrate that this dogma is misguided. We believe that many of the problems people have putting their knowledge of logic to work, whether in machines or in their own lives, stems from the logocentricity that has pervaded its study for the past hundred years. Recently, some researchers outside the logic tradition have explored uses of diagrams in knowledge representation and automated reasoning, finding inspiration in the work of Euler, Venn, and especially C. S. Peirce. This volume is a testament to this resurgence of interest in nonlinguistic representations in reasoning. While we applaud this resurgence, the aim of this chapter is to strike a cautionary note or two. Enchanted by the potential of nonlinguistic representations, it is all too easy to overreact and so to repeat the errors of the past.